Before I clone anything, I’ve started doing the unglamorous thing: I open the GitHub API, I count the files myself, and I read the code instead of the pitch. Half the time the number is inflated. The other half it’s real but it’s measuring the wrong thing.
This edition is three repos that all lead with a big number. One of them earns it in a way nobody’s talking about. One of them is quietly arguing with its own README. And one of them isn’t about the number at all — it’s about who gets to host the software now. The question, as always, isn’t “is this impressive?” It’s the only question that matters from inside a studio: does it change how I ship?
Let’s count.
1. agency-agents — an org chart you can install
msitarzewski/agency-agents — 107,289 ★, MIT, created October 2025. The caption that surfaced it said “147 specialized sub-agents.”
Here’s the honest accounting. The README claims 147 agents in one place and 144 in another — the number floats even for the author. When I pulled the file tree and counted the actual agent definitions, I got 235 markdown files across roughly sixteen “divisions”: Specialized (48), Marketing (36), Engineering (32), Game Development (20), Strategy (16), Integrations (15), Design (9), down to Spatial Computing (6). So the headline number isn’t inflated — if anything it undersells. But that’s also not the point, and treating it as a leaderboard score is exactly the mistake.
Because these aren’t 235 tools. They’re 235 personalities with a process. I opened spatial-computing/visionos-spatial-engineer.md and it isn’t a generic “you are a helpful assistant” stub. It declares a vibe (“Builds native volumetric interfaces and Liquid Glass experiences for visionOS”), then specs real platform knowledge: glassBackgroundEffect with configurable display modes, spatial widgets that snap to walls with persistent placement, RealityKit–SwiftUI observable entities. That’s not a prompt. That’s a job description.
Why a designer should care: two of these divisions land squarely in my world. Spatial Computing — visionOS, XR cockpit interaction, Metal rendering — and Game Development, with twenty agents that include narrative-designer.md, level-designer.md, and technical-artist.md alongside the Unity and Unreal specialists. If you run a studio that touches XR or gaming, this is “design = orchestration” made embarrassingly literal: not one overloaded agent, but an org chart that works in parallel.
The honest part: the value here is the structure, not the count. The real cost of adopting this isn’t installation — it’s the discipline to pick the four divisions you actually need and delete the other twelve. Install all 235 and you haven’t built an agency, you’ve built noise with a payroll. The number “147” is marketing. The org chart is the product.
2. ruflo — the swarm that argues with its own README
ruvnet/ruflo — 57,743 ★, MIT, by rUv (formerly Claude Flow, now powered by a Rust engine). The clip that flagged it said “60+ agents” and “it gets smarter every run.”
The clip undercounts. The README’s own comparison table claims 98 agents, 60+ commands, 30 skills, MCP server, hooks, daemon — the “60+” from the caption is actually the command count, not the agents. When I counted the agent definitions in .claude/agents/ I found 108 markdown files, organized into genuinely interesting clusters: a consensus/ folder with a byzantine-coordinator, a raft-manager, a gossip-coordinator. There is a real queen-coordinator.ts in the swarm source, not just a metaphor in the marketing. The hierarchy the clip described — Queen agents managing tactical workers — exists in the code.
But watch the tool count, because this is where the repo argues with itself. The headline says 314 MCP tools. Scroll down and a feature row quietly says “~210 tools, ready to call.” Same README, two numbers, a hundred tools apart. That’s not a lie — it’s the honest residue of a project moving faster than its own docs. Worth knowing before you build your week on the larger figure.
Why a designer should care: the thesis here isn’t “more agents.” It’s the router. Ruflo puts an adaptive layer between you and the swarm — simple tasks stay lightweight, complex ones spin up the full hierarchy. That’s orchestration turned into infrastructure, which is exactly the shift I keep circling: the designer’s job moves from making the artifact to designing the flow that makes it. This is the natural sequel to last edition’s argument — your agent doesn’t need a better model, it needs a better memory — except here the memory becomes coordination that supposedly composes across runs.
The honest part: “self-learning swarm intelligence” is the most inflatable phrase in the catalog right now. To ruflo’s credit, the README actually shows its homework — it links a benchmark and reports the vector memory is “~1.9x faster at N=20k” with a measured crossover where it ties or loses at small N. That’s the rarest thing in this whole report: a self-improvement claim with a caveat attached. But the real install cost is brutal honesty about the surface area. The README’s own pitch — “you don’t need to learn 314 MCP tools” — is a tell. When the onboarding has to reassure you about the size, the size is the cost.
3. odysseus — not a number, a verdict on who hosts the software
pewdiepie-archdaemon/odysseus — 45,025 ★, MIT, created at the end of May 2026 and already past 45k stars and 5,000 forks in days. This one’s big number is the growth curve. But the curve isn’t the story.
It’s a self-hosted AI workspace — the author’s words: “the self-hosted version of the UI experience you get from ChatGPT and Claude. But with more jank and fun.” Local-first, privacy-first, runs on your own hardware. And here’s the thing a designer should sit with: it isn’t invented, it’s assembled. The Agent mode is built on opencode. The Cookbook is built on llmfit. Deep Research is adapted from Alibaba’s Tongyi. Memory rides on ChromaDB. The repo even ships the upstream MIT licenses for opencode and llmfit in a licenses/ folder. This is “design = orchestration” applied to personal infrastructure: the craft is in the assembly and the seams, not the parts.
The detail I keep coming back to is the Cookbook. There are over a dozen cookbook files in the tree — cookbook-hwfit.js, cookbookServe.js, a whole CLI under scripts/odysseus-cookbook. What it does is real UX on a genuinely ugly problem: it scans your hardware, scores which models actually fit your VRAM, and downloads-and-serves with one click. The question every local-AI tourist asks — “what can my machine actually run?” — has lived in the terminal until now. Odysseus put a face on it.
And there’s a tell in start-macos.sh that I love, because it’s a design decision disguised as a shell script. The comment explains why the Mac install is native instead of Docker: “Docker on macOS is a Linux VM with no access to the Metal GPU. Running natively lets Cookbook detect and use your Mac’s GPU.” Someone made the harder packaging choice so the GPU would actually work. Launches at 127.0.0.1:7860.
Why a designer should care: when a creator with a hundred-million-plus audience spends his cultural capital on “keep your data on your own machine,” data sovereignty stops being a hobbyist niche and becomes a mainstream aesthetic. For anyone designing products, that’s a demand signal worth more than the star count. And the declared “jank and fun” is the actual thesis — it’s the anti-committee software, the thing that didn’t have to please a review board. A useful reminder of what gets sanded off when everything ships through corporate approval.
The honest part: self-hosted is not free. The star count is free; the hardware isn’t. Serving anything beyond a small model means real VRAM, and the Cookbook’s honesty is the point — on a 16GB Mac mini like mine, it’ll happily run the UI and small models, then tell you the truth about the big ones. “Works out of the box” holds right up until you want a model your box can’t hold. The maintenance is yours too. Sovereignty has an electricity bill.
The thread
Three repos, three headline numbers, and the numbers ranked exactly backwards from how much they’ll change your week. agency-agents undersold its real count and it barely matters, because the value was the org chart, not the headcount. ruflo’s count is real and so is the router, but the README can’t decide if it’s 314 tools or 210, and the surface area is the cost. odysseus’s number is the smallest claim it makes; the actual payload is a verdict on where software gets to live now.
The caption is the receipt. The code is the meal. Count the files yourself.
See you in the next Repo Report.